Multi-task transformer network for subject-independent iEEG seizure detection

• Published in: Expert systems with applications Vol. 268; p. 126282
• Main Authors: Sun, Yulin, Cheng, Longlong, Si, Xiaopeng, He, Runnan, Pereira, Tania, Pang, Meijun, Zhang, Kuo, Song, Xin, Ming, Dong, Liu, Xiuyun
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 05.04.2025
• Subjects: Channel-wise mixup; Intracranial EEG; Multi-task learning; Seizure detection; Subject-independent; Transformer network; Transformer network; Intracranial EEG; Channel-wise mixup; Seizure detection; Multi-task learning; Subject-independent
• ISSN: 0957-4174
• DOI: 10.1016/j.eswa.2024.126282

Subject-independent seizure detection algorithms are typically grounded in scalp electroencephalogram (EEG) databases, due to standardized channels and locations of EEG electrodes. Intracranial EEG (iEEG) has the characteristics of low noise and high temporal resolution compared with scalp EEG. However, it is still a big challenge for seizure detection using iEEG, because of the inconsistent number and locations of implanted electrodes in different patients, which results in a lack of unified algorithms. This study introduces an innovative approach for subject-independent seizure detection using iEEG, combining channel-wise mixup, transformer networks, and multi-task learning. Channel-wise mixup enhances data utilization by effectively leveraging information from different subjects, while multi-task learning improves the generalization of the model by concurrently optimizing both the seizure detection and the subject recognition tasks. 2983 files from two well-known epilepsy databases, i.e. SWEC-ETHZ and HUP were used in our study and the result showed that our approach surpasses currently existing methods. In terms of accuracy and generalization of seizure detection, our method achieved an area under the receiver operating characteristic curve (AUC) of 0.97 and 0.95 on the two databases respectively, which are significantly higher than the result of the currently existing methods. This study proposed a new method with great potential for surgery planning of epilepsy patients. •Transformer networks used for detecting seizures in patients with variable electrodes.•Multi-task learning improves performance by addressing inter-subject variability.•Channel-wise mixup method augments the ictal period data by introducing diversity.•Experiments on 75 subjects from 2 databases confirm the effectiveness of the method.